How many LLM users are exposed to Russian propaganda?
After theorizing LLM grooming last year, we estimate that users are exposed to hundreds of thousands of Russia-influenced chatbot responses per month.
Our information environment has been drastically reshaped in recent years with the advent of LLMs. Now boasting more than a billion users globally, tools like ChatGPT, Perplexity, and Claude have become a primary news source for many people, research tools for others, and time-saving task managers. There is no doubt: LLMs are ubiquitous. With new technologies, however, comes new attempts to manipulate them for nefarious purposes. We’ve seen this before; for example, when household internet usage skyrocketed in the 2000s, phishing and spoofed websites barreled onto the scene as well.
The American Sunlight Project first published our report theorizing LLM Grooming — the mass-production and duplication of false narratives online with the intent of manipulating LLM outputs — last February. In it we analyzed the Pravda Network, a centralized collection of websites that spread pro-Russian propaganda — to the tune for 3+ million articles per year — in dozens of languages, aimed at countries ranging from Niger to Japan to Serbia to Canada. We found, at the time, the network’s user interface so hostile to human users that we thought automated projection into the digital ether appeared to be the network’s primary intent. The consequences have already been measurable — NewsGuard, the Institute for Strategic Dialogue, the Atlantic Council, and the European Digital Media Observatory have since engaged in critical research documenting instances of LLM Grooming.
Yet a critical question has remained unaddressed: At what scale are ordinary chatbot users actually being exposed to groomed outputs? This post presents new research in which we attempt to quantify that exposure directly in the European Union (EU), moving the conversation from vulnerability in principle to harm in practice.
We estimate that chatbot users across the European Union interact with at least 500,000 Russian-influenced, Ukraine-related responses per month.
Finding this number required the construction of a novel data pipeline, but not in the coding-kind-of-way that implies. Rather, we searched for four other numbers, each acting as a subset of the former:
Number of news-related chatbot visits/month
Number of Ukrainian news-related chatbot visits/month
Number of Ukrainian news-related chatbot responses/month
Number of Russian-influenced Ukrainian-related chatbot responses/month
Four steps to get the results:
Step 1: Number of news-related chatbot visits/month: First, we found the percentage of people who used chatbots for news in each EU country using the 2025 Reuters Digital News Report. We multiplied each country’s percentage by their population. This gave us a rough number of news-related chatbot visits each month in each EU country.
Step 2: Number of Ukrainian news-related chatbot visits/month: Next, we estimated how popular Ukraine was as a news topic using Google Trends. To illustrate how Google Trends works, let’s imagine there are only two things people search on Google: “Trump” and “FIFA World Cup.” Google Trends lets us track how popular Trump is compared to the World Cup. The most popular topic, in this case the World Cup, always gets a score of 100. Trump’s score is based on how much less he is searched than the World Cup. If Trump is searched half as much as the World Cup, Trump will have a score of 50 and the World Cup will have a score of 100. Therefore, “Trump” would be 33% (50/150) of all Google searches. Essentially, we multiplied each country’s percentage of Ukraine news-related searches by their population.
Step 3: Number of Ukrainian news-related chatbot responses/month: After finding the number of Ukrainian news-related visits, we still needed to convert the number of visits to the number of responses. To find this, we used a Neilsan Norman Group study that found the average number of user prompts per visit to major chatbots. This approach assumes that the number of user prompts per visit regarding Ukraine news is equal to the average number of user prompts per visit on any topic, which we believe likely creates a conservative underestimate, as complex topics like news are likely to lead to more prompts than average.
Step 4: Number of Russian-influenced, Ukrainian-related chatbot responses/month: Finally, we found the number of Ukrainian-related chatbot responses that were Russian-influenced using the Institute for Strategic Dialogues’ latest research, which queried various chatbots about the Russian conflict in Ukraine to determine the rate of Russian-influenced responses (roughly 18%). This gave us our final result: the number of Russian-influenced Ukrainian-related chatbot responses per month.
Our calculations illustrating these steps and estimates of exposure per country can be viewed here. In total, we estimate chatbot users in the EU are exposed to at least 504,060 Russian-influenced, Ukraine-related chatbot responses per month.
Limitations and implications
The most appropriate interpretation of the results would be to view them as a conservative baseline. There are several potential underestimations we may have made, for example:
By multiplying the percentage of people who visit chatbots for news by the population of a country, to find the number of visits in a given country, the result essentially assumes each person who visits chatbots for news only visits once per month. In reality, these individuals likely visit chatbots more than once per month.
Queries that included Ukraine, and a percentage of those that included Russia, were the only ones counted toward the popularity score of Ukraine as a topic. However, it is inevitable that Ukrainian-adjacent queries that contained neither of those words, such as “Trump” or “war,” would likely have some intent to acquire information about Ukraine.
News-related queries may yield a greater number of prompts per conversation than other types of prompts.
However, even if the estimation of these results is not exact, they should serve as a warning sign. By scaling up, for example, Germany’s results to per-year, one can better understand the scale: 1,520,520 Russian-influenced Ukrainian-related chatbot responses (126,710 per month). Now, let’s say the German public’s use of chatbots as news sources doubled over five years. With all else equal, the result is 3,041,040 responses per year, however doubling this percentage still leaves it at only 8% of the population. With the rapid pace of technological change, it is not unreasonable to fathom a medium-term future where, say, 32% of Germans use chatbots as a news source. In that case, the number becomes 12,164,160 Russian-influenced responses per year.
The rabbit hole can go deeper. For instance, with further expansion of the Pravda network or other networks, Russia might increase its share of sources cited by chatbots about the invasion of Ukraine from 18%. Or, more people in the EU might begin to use chatbots for information about the full-scale invasion — or future international conflicts started by the Kremlin. This does not even mention the secondary effect of news outlets using these responses as a source themselves, which has its own dangerous implications.
The most important finding this analysis confirms is that Russia has its foot in the chatbot door and there are many ways it can force it open wider. In the wake of the U.S. government’s pullback from both promoting information literacy and combatting overt foreign influence — vital, nonpartisan work which protects democracy — the continued effects of LLM Grooming pose a real threat to the safety and security of elections, governing institutions, and real, human lives around the globe.
 | A guest post by
|